I built a Chrome extension

I built a Chrome extension that extracts code from tutorial
videos using Claude's Vision API. Wrote up what I learned:

Key findings:

1. Vision models handle code extraction surprisingly well better than OCR for anything syntax-sensitive

2. Prompt engineering matters more than model choice.
   "Extract the code" vs "identify all code blocks, preserve
   indentation, differentiate from terminal output" huge
   difference in quality

3. Latency: ~2-3 seconds for a 1080p frame. Acceptable for
   this use case, too slow for real-time

4. Edge case: slides with code + screenshots of code —
   model correctly identifies both as "code" even with
   different visual styling

Full writeup with prompt examples: [link to dev.to article]

The extension itself is free if anyone wants to try the
output quality: codera.click